fix: comprehensive bugfixes — 53 issues resolved#2
Conversation
Critical fixes: - ws_client.py: absolute import → relative import - server.py: PID file encoding + write after bind (race condition) - client.py: port-check loop after daemon restart (race condition) - spawn.py: temp file cleanup via atexit (resource leak) - server.py: localhost-only check for /api/shutdown (DoS prevention) High fixes: - helpers.py: MD5 → SHA-256 for content dedup - client.py: log rotation (1MB limit) - ops_ai.py: validate human_secret before HMAC - ops_chat.py: log WebSocket failures instead of silent pass - crypto.py: catch InvalidToken specifically - db.py: document WAL read safety - ops_room.py: int() validation for --port - server.py: 100KB per-message content limit - ops_peer.py: consistent peer ID generation Medium fixes: - __init__.py: return exit code from output() (testability) - ops_chat.py: int() validation for -n flag - ops_ai.py: int() validation for --timeout/--rounds - spawn.py: atexit for temp file cleanup - relay_client.py: always send peer_name/peer_tool - db.py: move SELECT inside lock - ops_room.py: relay warning → info level - constants.py: fallback chain for get_machine_id() - ops_chat.py: join poller thread on exit - helpers.py: getpass.getuser() fallback for identity - client.py: consistent dict return from _req() Low fixes: - ws_client.py: socket timeout + ping/pong keepalive - ops_chat.py: copy message before mutating - ops_ai.py: dynamic @mention regex from TOOL_COMMANDS - ops_room.py: --name flag for discuss command - ops_ai.py: pass context in quick invites - conftest.py: try HTTP shutdown before taskkill /F - test_db.py: close DB before temp dir cleanup Documentation: - PRD.md: fix argparse reference → manual dispatch - PRD.md: update test count 40 → 69 - SKILL.md: clarify 16 sub-commands vs 12 top-level commands
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 21eb9bf75f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| cmd = [] | ||
| for part in config["cmd"]: | ||
| if part == "{prompt}": | ||
| cmd.extend(["--file", _tmp_prompt_file]) |
There was a problem hiding this comment.
On Windows, any non-stdin tool with a prompt over 32 KB takes this branch, so codex is invoked as codex exec --full-auto --file <tmp>. The official Codex CLI reference documents PROMPT as string | - (read stdin) for codex exec, not a --file flag (https://developers.openai.com/codex/cli/reference#codex-exec). In Windows rooms with enough history/context, inviting Codex will fail before it can respond; restrict this fallback to tools that actually support file attachments or use stdin (-) for Codex.
Useful? React with 👍 / 👎.
| pid = make_ai_peer_id(tool) | ||
| return client.register_peer(pid, name, ptype, tool, get_machine_id(), host, port) |
There was a problem hiding this comment.
For the documented manual remote registration flow (peer register ... --host 192.168.1.50), this now generates the ID from the local machine and stores the local machine in the machine field. Registering a remote Codex peer therefore collides with/overwrites the local Codex peer or any other remote Codex peer instead of creating a host-specific peer, so peer list loses the remote endpoint metadata needed to distinguish participants.
Useful? React with 👍 / 👎.
| if len(b["content"]) > 100 * 1024: # 100KB per-message limit | ||
| return self._json({"error": "message content too large (max 100KB)"}, 413) |
There was a problem hiding this comment.
When a public-room message hits this new 413 path, _post_message treats the returned {'error': ...} as a truthy local message and still writes the content to the relay without a local message id. For messages or AI responses over 100KB, the CLI can report success while SQLite never stores the message (and local-only sends also return ok: true with an embedded error), so local history/export diverges from the relay. Raise on the local rejection before the relay write or enforce this limit before returning success.
Useful? React with 👍 / 👎.
| except socket.timeout: | ||
| # Idle timeout — send ping to keep connection alive | ||
| self._send_ping(b"") | ||
| continue |
There was a problem hiding this comment.
On an idle timeout this loop sends a ping and immediately continues, while _send_ping swallows send failures and no pong is required before keeping ws_connected true. In interactive relay chats where the network path becomes half-open or the relay stops responding without closing the TCP socket, the poller keeps skipping HTTP polling indefinitely, so new messages are not displayed until the process is restarted. Treat a missed ping/pong or send failure as a disconnect so the existing polling fallback can run.
Useful? React with 👍 / 👎.
| try: | ||
| import getpass | ||
| username = getpass.getuser() |
There was a problem hiding this comment.
When neither USER nor USERNAME is set, previous versions generated the human id with the stable fallback name user; this new fallback commonly returns the OS account such as root. Upgrading in that environment rewrites an existing *-human-user identity to *-human-root, so prior room membership and relay identity no longer match the user's existing peer. Preserve the old fallback for existing identities or only use getpass when creating a brand-new identity.
Useful? React with 👍 / 👎.
1 participant
Summary
Comprehensive bugfix addressing 53 identified issues across the AI Peer codebase. All changes are backwards-compatible and use only Python stdlib.
Critical Fixes (5)
High Fixes (9)
Medium Fixes (12)
Low Fixes (16)
Documentation (3)
Test Fixes
Test Results
All 66 tests passing on Windows with Python 3.10+.
Changes
18 files changed, 247 insertions(+), 68 deletions(-)